[RFC/PATCH] epoll: replace EPOLL_CTL_DISABLE with EPOLL_CTL_POKE

[Eric Wong]

"Paton J. Lewis" <palewis@xxxxxxxxx> wrote:
> --- /dev/null
> +++ b/tools/testing/selftests/epoll/test_epoll.c

> +	/* Handle events until we encounter an error or this thread's 'stop'
> +	   condition is set: */
> +	while (1) {
> +		int result = epoll_wait(thread_data->epoll_set,
> +					&event_data,
> +					1,	/* Number of desired events */
> +					1000);  /* Timeout in ms */

<snip>

> +		/* We need the mutex here because checking for the stop
> +		   condition and re-enabling the epoll item need to be done
> +		   together as one atomic operation when EPOLL_CTL_DISABLE is
> +		   available: */
> +		item_data = (struct epoll_item_private *)event_data.data.ptr;
> +		pthread_mutex_lock(&item_data->mutex);

The per-item mutex bothers me.  Using EPOLLONESHOT normally frees me
from the need to worry about concurrent access to the item userspace.

Instead of having another thread call delete_item() on a successful
EPOLL_CTL_DISABLE, I'd normally use shutdown() (which causes
epoll_wait() to return the item and my normal error handling will kick
in once I try a read()/write()).

However, since shutdown() is limited to sockets and is irreversible,
I came up with EPOLL_CTL_POKE instead.  Comments greatly appreciated
(I'm not a regular kernel hacker, either)